Alexandre Eid defends his PhD on March 17, 2022.
Place : UCBL, Bât. Berthollet (Villeurbanne).
Note : by invitation only, please contact Prof. Guy Clerc.
Jury :
Rapporteurs : Mohamed Benbouzid, Mitra Fouladirad
Examinateur : Marie-Cécile Péra, Nicolas Bonneel, Laetitia Chapel, Hubert Razik
Encadrement : Guy Clerc
Invités : Badr Mansouri, Babak Nahid-Mobarakeh, Noureddine Takorabet
Abstract :
This thesis develops a complete predictive maintenance method from data pre-processing to the determination of the remaining lifetime of an electric thrust reverser. In doing so, its objective is to provide a starting point for implementing a more global strategy of industrial predictive maintenance. After defining the industrial context of this work, we will present the state of the art on methods of PHM (Prognostics & Health Management). The methodology aims at solving two types of problems: classification or partitioning to perform a diagnostic step and regression coupled with classification to perform the prognosis. Since the last few years, artificial intelligence methods and, in particular, deep neural networks have made it possible to provide a more relevant solution to these two major classes of problems. That is why we will use artificial intelligence algorithms to develop the whole method. From vibratory signals measured on the thrust-reverser are then extracted descriptors of defects. These are statistical quantities whose dynamics are expected to be correlated with the system’s aging. However, these descriptors do not have a label allowing to assign a degree of severity to a given point. In other words, the state of health is unknown over time. Therefore, a time series partitioning algorithm is developed to perform this task. Each descriptor is encoded as an image segmented by a deep neural network trained by artificial data generation. One-dimensional information on the group boundaries is then extracted and filtered. Then, a kernel density estimation transforms the resulting signal into an empirical probability density. Finally, a Gaussian mixture model extracts the independent components of the created distribution to obtain probable group boundaries. This method, SUNRISE (Soft Clustering for Time Series), allows revealing different degrees of severity of faults in the studied data, with their respective likelihood, without any prior knowledge of the data structure. Moreover, to evaluate the obtained partitioning results, we have developed a new quality indicator. It measures the temporal consistency and the similarities of the group shape obtained by the partitioning scheme. Once labeled data are possessed, the determination of the remaining lifetime of the system is performed by time series alignment using the method PARTITA-RULE (Partial Time scaling Invariant Temporal Alignment for Remaining Useful Life Estimation). The idea is to determine the health status of a new actuator by measuring the similarity between its vibration signals and the previous ones contained in the database. A time series alignment method is developed for this: PARTITA. Once the unknown series is aligned with all possible candidates in the database, a weighting scheme is created to assign a likelihood score to each candidate. Finally, a fusion model solving an optimal transport problem is used to obtain the remaining lifetime or RUL of the actuator with an uncertainty measure. Subsequently, the robustness of the method to incomplete time segments is tested. This thesis resulted in a method of PHM created specifically to operate in an industrial setting with a small input dataset but which could be enriched as the product is used. Finally, the method is robust to missing data and remains promising under multiple operational conditions.
Keywords :
prognosis; aeronautics; electromechanical actuators; clustering; time series; sequence alignment; deep learning; optimal transport;